Game addiction

    Studying data from Steam

    by Sebastian Wolf

    Task

    Your client is a mental health expert from an NGO who is interested in understanding more about gaming and the potentially addictive effect it can have on some individuals. You are meeting the client in a few days and they would like you to extract and present insights from the Steam dataset to help them in their research.

    Please use whichever tools you feel the most comfortable with, but we do recommend Tableau which is a popular choice. Tableau is free for students and there is also a free trial available here.

    My approach in these slides

    • Explore the context
      • Geographic distribution of gamers
      • Evolution of gamers on Steam
      • Times when gamers are active
    • Analyze game addiction
      • Define game addiction in the dataset
      • Explore correlations of playing time with several engagement factors
      • Derive conclusions and suggest focus for further analysis

    Setup

    In [1]:
    # Make sure script changes take effect within this session
%load_ext autoreload
%autoreload 2
    In [2]:
    # import some useful packages for this analysis, start spark session
from setup import *
    In [3]:
    spark
    Out[3]:

    SparkSession - in-memory

    SparkContext

    Spark UI

    Version
    v2.4.5
    Master
    local[*]
    AppName
    pyspark-shell

    Open localhost:4040 to monitor the spark UI

    Load csvs

    In [4]:
    # list of data we want to load
dfs = ['Player_Summaries', 'Games_Publishers', 'Games_Genres', 'Games_Developers', 'Games_1', 'Games_2', 'Friends', 'Groups', 'App_ID']
    In [5]:
    # load raw data and count rows
spark_handler = spark_df_handler()
print('Number of records \n')
for each in dfs:
    spark_handler.load(each)
    print('{0} : {1}'.format(each, str(spark_handler.dfraw[each].count())))

    Number of records 

Player_Summaries : 5000000
Games_Publishers : 18761
Games_Genres : 39669
Games_Developers : 19333
Games_1 : 15426029
Games_2 : 29711963
Friends : 16450558
Groups : 3383666
App_ID : 17783

    Prepare tables

    Tables centred on the player

    Player summaries

    In [6]:
    # Create timestamp vars and select vars of interest
player_summaries = spark_handler.dfraw[dfs[0]]
player_summaries = player_summaries.withColumn('datetimecreated', F.to_timestamp(player_summaries.timecreated,'yyyy-MM-dd HH:mm:ss'))
player = player_summaries.select('steamid', 'primaryclanid', 'datetimecreated')
    In [7]:
    # make a seperate df with additional info on signup times
player_signup = player.withColumn('year_created', F.date_trunc('year', player_summaries['datetimecreated']))
player_signup = player_signup.withColumn('month_created', F.date_trunc('month', player_signup['datetimecreated']))
player_signup = player_signup.withColumn('day_created', F.date_trunc('day', player_signup['datetimecreated']))
player_signup = player_signup.withColumn('dayofweek_created', F.dayofweek('datetimecreated').cast('byte'))
player_signup = player_signup.withColumn('dayhour_created', F.date_trunc('hour', player_signup['datetimecreated']))
player_signup = player_signup.withColumn('hour_created', F.hour('datetimecreated').cast('byte'))

    Player stats

    In [8]:
    # concatenate game tables
games = spark_handler.dfraw[dfs[4]] #.union(spark_handler.dfraw[dfs[5]])
    In [9]:
    # load app table
app_info = spark_handler.dfraw[dfs[8]]
    In [10]:
    # join games and appinfo
games = games.join(app_info, 'appid', how = 'left')
    In [11]:
    # calculate player stats using the game table
player_stats = games.groupby('steamid').agg(F.sum('playtime_2weeks').alias('playtime_2weeks_player'),
                                            F.sum('playtime_forever').alias('playtime_forever_player'),
                                            F.countDistinct('appid').alias('game_count_player'),
                                            F.sum('Is_Multiplayer').alias('multiplayer_count_player'),
                                            F.sum('Price').alias('spending_player'))

# fill in zeros for players that didn't play in the two weeks prior, and convert times to hours
player_stats = player_stats.fillna(0)
player_stats = player_stats.withColumn('playtime_average_player', F.col('playtime_2weeks_player') / (14 * 60))
player_stats = player_stats.drop('playtime_2weeks_player')

# calcuate fraction of multiplayer games
player_stats = player_stats.withColumn('multiplayer_fraction_player', F.col('multiplayer_count_player') / F.col('game_count_player'))
    In [12]:
    # Save player stats
# player_stats_pd = player_stats.toPandas()
# player_stats_pd.to_csv(os.path.join(path, '../player_stats.csv'))
    In [13]:
    # player_stats_pd = pd.read_csv(os.path.join(path, '../player_stats.csv'), index_col = False)

    Group stats

    In [14]:
    # load group table
groups = spark_handler.dfraw[dfs[7]]
    In [15]:
    # count number of groups per player
group_stats = groups.groupby('steamid').agg(F.countDistinct('groupid').alias('group_count_player'))

    Friends stats

    In [16]:
    # load friends table and create some vars
friends = spark_handler.dfraw[dfs[6]]
friends = friends.withColumn('friend_since', F.to_timestamp(friends.friend_since,'yyyy-MM-dd HH:mm:ss'))
friends = friends.withColumn('friend_since_day', F.date_trunc('day', friends['friend_since']))
friends = friends.withColumn('friend_since_month', F.date_trunc('month', friends['friend_since']))
    In [17]:
    # join friends table with player stats
friends_stats = friends.join(player_stats, friends.steamid_b == player_stats.steamid, how = 'left')
    In [18]:
    # calculate friends' stats
agg_friends_stats = friends_stats.groupby('steamid_a').agg(F.countDistinct('steamid_b').alias('friend_count'),
                                                           F.mean('playtime_average_player').alias('playtime_average_friends'),
                                                           F.max('playtime_average_player').alias('playtime_average_friends_max'),
                                                           F.mean('playtime_forever_player').alias('playtime_forever_friends'),
                                                           F.max('playtime_forever_player').alias('playtime_forever_friends_max'),
                                                           F.mean('game_count_player').alias('game_count_friends'),
                                                           F.mean('multiplayer_count_player').alias('multiplayer_count_friends'),
                                                           F.mean('multiplayer_fraction_player').alias('multiplayer_fraction_friends'),
                                                           F.mean('spending_player').alias('spending_friends'))

    Join player summaries, player stats, group stats, friends stats

    In [19]:
    # Join it all together
user_summary = player_stats.join(player, 'steamid', how = 'left')
user_summary = user_summary.join(group_stats, 'steamid', how = 'left')
user_summary = user_summary.join(agg_friends_stats, user_summary.steamid == agg_friends_stats.steamid_a, how = 'left')
    In [20]:
    # Save output
user_summary_pd = user_summary.toPandas()
user_summary_pd.to_csv(os.path.join(path, '../user_summary.csv'))
    In [4]:
    # Read output
user_summary_pd = pd.read_csv(os.path.join(path, '../user_summary.csv'), index_col = False)
    In [22]:
    # Check sanity
user_summary_pd.describe()
    Out[22]:
    Unnamed: 0 steamid playtime_forever_player game_count_player multiplayer_count_player spending_player playtime_average_player multiplayer_fraction_player group_count_player steamid_a friend_count playtime_average_friends playtime_average_friends_max playtime_forever_friends playtime_forever_friends_max game_count_friends multiplayer_count_friends multiplayer_fraction_friends spending_friends
    count 1,665,547 1,665,547 1,665,547 1,665,547 1,665,547 1,665,547 1,665,547 1,665,547 431,952 873,114 873,114 217,836 217,836 217,836 217,836 217,836 217,836 217,836 217,836
    mean 832,773 76,561,198,029,079,680 13,894 9 5 118 0 1 7 76,561,198,028,959,792 14 1 1 50,350 65,973 25 12 1 320
    std 480,802 34,592,379 31,703 24 9 303 1 0 16 35,733,490 25 2 2 53,796 75,006 45 17 0 544
    min 0 76,561,197,960,265,856 0 1 0 0 0 0 1 76,561,197,960,265,856 1 0 0 0 0 1 0 0 0
    25% 416,386 76,561,197,998,434,720 144 1 1 15 0 0 1 76,561,197,997,833,664 2 0 0 11,671 12,830 4 3 0 45
    50% 832,773 76,561,198,027,902,160 1,862 3 2 35 0 1 2 76,561,198,027,900,560 5 0 0 35,657 42,696 10 7 1 121
    75% 1,249,160 76,561,198,065,610,304 12,600 7 5 85 0 1 6 76,561,198,067,173,264 14 1 2 71,351 93,006 26 14 1 365
    max 1,665,546 76,561,198,086,379,456 1,709,165 1,528 464 14,973 89 1 1,511 76,561,198,086,366,912 674 25 25 1,709,165 1,709,165 1,474 421 1 14,413

    Exploring the context

    • Geographic distribution of gamers
    • Evolution of gamers on steam
    • When gamers are active

    The Steam dataset includes gamers from the entire world

    • Yet, most steam users report their location to be in the United States and Russia
    • Explore user counts per country by hovering over the map.
    In [23]:
    # get aggregates per country
player_location = player_summaries.groupby('loccountrycode').count().toPandas()
    In [24]:
    # load country borders
world_borders = gpd.read_file(os.path.join(path, '../geo_files/TM_WORLD_BORDERS-0.3.shp'))
    In [25]:
    # merge borders with agg stats
map_df = world_borders.merge(player_location, how = 'left', left_on = 'ISO2', right_on = 'loccountrycode')
    In [26]:
    # get density and per capita
map_df['player_density'] = map_df['count'] / map_df.AREA
map_df['player_per_cap'] = map_df['count'] / map_df.POP2005 * 100
    In [27]:
    # fill missing values
map_df = map_df.replace(np.inf, 0).fillna(0)
    In [28]:
    # map it out
m = folium.Map(location=[50, 0],
    tiles='CartoDB positron',
    zoom_start=1.5,
    control_scale = True)

colorscale_t = branca.colormap.linear.YlOrRd_09.scale(0, 100000)
colorscale_t.caption = '# of players registered'

folium.GeoJson(map_df, name='players per country', 
    style_function=lambda x: {"weight": 0.2, 'color': 'black', 'fillColor':colorscale_t(x['properties']['count']), 'fillOpacity':0.9},
    highlight_function=lambda x: {'weight':1, 'color':'black'},
    smooth_factor=1.0,
    show = True,
    tooltip=folium.features.GeoJsonTooltip(fields = ['NAME','count', 'player_density', 'player_per_cap'],
                                              aliases=['Country','Number of Players', 'Players / 10 sqkm', 'Players / 100 people' ], 
                                              labels=True, 
                                              sticky=True,
                                              localize=True
                                             )).add_to(m)
folium.LayerControl().add_to(m)
colorscale_t.add_to(m, 'scale')
m.save(os.path.join(git, '../html_maps/players_per_country.html'))

m
    Out[28]:

    Steam gaming is social

    • The number of new friendships is proportional to the number of new users
    • We see 2-3 times as many new friendships as new users per day
    In [29]:
    # count new players per day
new_player_count = player_signup.groupby('month_created').count().toPandas().set_index('month_created').sort_index()
    In [49]:
    # count new friends per day, ever since they started recording them
new_friendship_count = friends.where((F.col('friend_since_month') > dt.datetime(2009,1,1)) & \
                    (F.col('friend_since_month') < dt.datetime(2013,4,1))).groupby('friend_since_month').count().toPandas().set_index('friend_since_month').sort_index()
    In [50]:
    # increase font sizes for plots
plt.rcParams.update({'axes.titlesize': 25,
                     'axes.labelsize' : 20,
                     'legend.fontsize' : 20,
                     'xtick.labelsize' : 18,
                     'ytick.labelsize' : 18,})
    In [51]:
    # Summary plots by day

# Set up frame
fig, axes = plt.subplots(nrows=1,ncols=1, figsize = (20, 10))
fig.subplots_adjust(hspace = 0.4)

# plot
sns.lineplot(x = new_player_count.index, y = new_player_count.values[:,0], ax = axes, label = 'New accounts per day',  linewidth = 2)
sns.lineplot(x = new_friendship_count.index, y = new_friendship_count.values[:,0], ax = axes, label = 'New friendships per day (starting 2009)',  linewidth = 2)


axes.set(ylabel = 'Count', xlabel = 'Year')
axes.xaxis.set_major_locator(mdates.YearLocator())
axes.xaxis.set_major_formatter(mdates.DateFormatter('%y'))
axes.xaxis.set_minor_locator(mdates.MonthLocator(interval=1))
axes.legend(loc='upper left')

plt.show()

    Gamers are more active on weekends and mornings

    • The figure shows the day of the week and the hour when gamers sign up for Steam
    • We see heightened activity from Saturday morning to Monday morning
    In [52]:
    # make a crosstab for signup times
dfs, style_dfs = crosstab_list_count(player_signup.where((F.col('dayofweek_created').isNull() == False) & (F.isnan('dayofweek_created') == False)),
                                      main_var = 'dayofweek_created',
                                      list_of_pivot_vars = ['hour_created'],
                                      title = "Total number of new accounts created by day of week and ")
    In [54]:
    # plot signup times per hour

# Set up frame
fig, axes = plt.subplots(nrows=1,ncols=1, figsize = (20, 10))

# plot
signup_times = dfs['hour_created'].transpose()
signup_times.columns = ['Monday', 'Tuesday', 'Wednesday', 'Thursday', 'Friday', 'Saturday', 'Sunday']
signup_times.plot(ax = axes, linewidth = 2, color = ['red', 'grey', 'grey', 'grey', 'grey','blue', 'green'])

axes.set(ylabel = '# of new users', xlabel = 'Hour of the day')
axes.legend(loc='upper left')

plt.show()
    In [55]:
    # dispaly the signup crosstab
style_dfs['hour_created']
    Out[55]:
    Total number of new accounts created by day of week and hour_created
    0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
    dayofweek_created
    1 20,712 24,111 26,832 31,342 34,095 36,218 38,641 40,540 42,115 43,042 43,233 42,606 39,839 35,783 31,119 26,809 22,926 20,648 19,307 17,880 16,454 14,946 14,116 13,566
    2 14,412 15,932 18,235 20,946 24,119 27,304 31,342 34,746 37,464 39,771 41,727 40,832 39,022 34,727 29,931 25,468 21,908 20,008 18,899 17,768 16,401 14,957 13,890 13,366
    3 14,313 16,409 18,892 21,698 24,832 28,003 32,019 34,880 37,742 39,369 41,241 41,093 38,825 34,088 29,350 25,058 21,991 20,085 19,911 18,395 17,032 15,379 14,445 14,331
    4 14,790 16,674 18,663 21,668 24,928 29,029 32,721 35,825 38,279 40,294 41,175 40,272 37,919 34,552 30,678 26,029 22,960 21,000 19,750 18,820 17,630 16,241 15,168 14,850
    5 15,722 17,339 19,468 22,754 25,457 28,868 32,343 35,709 38,432 41,012 42,140 42,238 40,476 36,321 32,086 27,216 23,704 21,386 20,126 18,942 17,688 16,515 15,676 15,540
    6 16,600 18,336 20,696 24,241 27,735 32,039 36,887 40,296 42,730 45,495 46,455 46,023 43,803 40,327 36,501 31,101 27,244 24,522 23,321 22,993 22,021 21,111 20,438 20,172
    7 21,960 24,851 29,457 33,788 37,520 40,902 43,537 46,180 48,280 50,307 51,070 48,617 45,448 41,444 36,958 31,822 27,636 24,587 22,707 22,155 21,191 20,189 19,576 19,505

    Analyzing game addiction

    • Define addiction for this analysis
    • Explore correlations with engagement factors in 5 hypothesis
    • Summarize findings
    • Derive conclusions and suggest focus for further analysis

    Game addiction affects only a small fraction of gamers

    • The median playtime among active gamers (users that were active in the 2 weeks prior to data retrieval) was only 30min per day
    • Yet, 25% of active gamers play nearly 2h per day
    • We will focus on the 1% of all gamers that play more than 5h per day on average
    • To learn about game addicition, we will explore correlations of playtime with several engagement factors
    In [5]:
    # Restrict to a subset of vars
user_summary_pd_subset = user_summary_pd.loc[:,['playtime_average_player', 
                                                'playtime_average_friends',
                                                'playtime_average_friends_max',
                                                'playtime_forever_player', 
                                                'playtime_forever_friends',
                                                'game_count_player', 
                                                'friend_count', 
                                                'group_count_player', 
                                                'multiplayer_fraction_player',
                                                'spending_player']]
    In [6]:
    # Remove implausible outliers
user_summary_pd_subset = user_summary_pd_subset[user_summary_pd_subset.playtime_average_player < 24]
    In [58]:
    # show some more decimals
pd.options.display.float_format = '{:,.4f}'.format
    In [59]:
    # check sanity
user_summary_pd_subset.describe()
    Out[59]:
    playtime_average_player playtime_average_friends playtime_average_friends_max playtime_forever_player playtime_forever_friends game_count_player friend_count group_count_player multiplayer_fraction_player spending_player
    count 1,665,521.0000 217,832.0000 217,832.0000 1,665,521.0000 217,832.0000 1,665,521.0000 873,104.0000 431,942.0000 1,665,521.0000 1,665,521.0000
    mean 0.2813 1.0057 1.4484 13,888.8125 50,349.3465 9.2606 13.7118 6.6285 0.6318 117.7021
    std 1.0252 1.6848 2.3496 31,654.5027 53,796.2542 24.4168 25.4832 15.9215 0.3582 302.7556
    min 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 1.0000 1.0000 0.0000 0.0000
    25% 0.0000 0.0000 0.0000 144.0000 11,670.0000 1.0000 2.0000 1.0000 0.4000 14.9900
    50% 0.0000 0.1607 0.2214 1,862.0000 35,656.1667 3.0000 5.0000 2.0000 0.6923 34.9700
    75% 0.0000 1.4131 2.1060 12,599.0000 71,350.2500 7.0000 14.0000 6.0000 1.0000 84.9400
    max 23.9952 24.9417 24.9417 1,709,165.0000 1,709,165.0000 1,528.0000 674.0000 1,511.0000 1.0000 14,972.5600
    In [60]:
    # Find 99th percentile
user_summary_pd_subset.playtime_average_player.quantile(0.99)
    Out[60]:
    5.113095238095238
    In [61]:
    # Playtime histogram

# Set up frame
fig, axes = plt.subplots(nrows=1,ncols=1, figsize = (20, 10))

# plot
user_summary_pd_subset.playtime_average_player.hist(bins = 24, ax = axes, linewidth = 2, color = 'g')
axes.set(ylabel = 'Count of users', xlabel = 'Hours played per day')
axes.text(7, 850000, '1 percent of (active) users play more than 5 hours per day on average', fontsize = 20)
axes.axvline(x = 5, color ='r')
patch = matplotlib.patches.Polygon(np.array([[5,0],[5,1700000],[26,1700000],[26,0]]), color = 'r', alpha = 0.1)
axes.add_patch(patch)
plt.show()
    In [7]:
    # Restrict to 99th percentile
user_summary_pd_subset_addicted = user_summary_pd_subset[user_summary_pd_subset.playtime_average_player > 5]
    In [63]:
    # correlations for heavy gamers
user_summary_pd_subset_addicted.corr()
    Out[63]:
    playtime_average_player playtime_average_friends playtime_average_friends_max playtime_forever_player playtime_forever_friends game_count_player friend_count group_count_player multiplayer_fraction_player spending_player
    playtime_average_player 1.0000 0.0311 0.0526 0.2559 0.0209 0.0010 0.0213 0.0168 -0.0553 -0.0046
    playtime_average_friends 0.0311 1.0000 0.8145 -0.0522 0.3686 -0.0976 0.0395 -0.0260 0.1347 -0.1098
    playtime_average_friends_max 0.0526 0.8145 1.0000 0.0582 0.3450 -0.0431 0.3303 0.0533 0.0863 -0.0546
    playtime_forever_player 0.2559 -0.0522 0.0582 1.0000 0.2668 0.2447 0.2968 0.1724 -0.1605 0.2478
    playtime_forever_friends 0.0209 0.3686 0.3450 0.2668 1.0000 0.0912 0.1706 0.0767 -0.0635 0.0894
    game_count_player 0.0010 -0.0976 -0.0431 0.2447 0.0912 1.0000 0.0853 0.0883 -0.3046 0.9870
    friend_count 0.0213 0.0395 0.3303 0.2968 0.1706 0.0853 1.0000 0.3090 0.0640 0.0782
    group_count_player 0.0168 -0.0260 0.0533 0.1724 0.0767 0.0883 0.3090 1.0000 -0.0392 0.0845
    multiplayer_fraction_player -0.0553 0.1347 0.0863 -0.1605 -0.0635 -0.3046 0.0640 -0.0392 1.0000 -0.3141
    spending_player -0.0046 -0.1098 -0.0546 0.2478 0.0894 0.9870 0.0782 0.0845 -0.3141 1.0000
    In [64]:
    # correlations for all gamers
user_summary_pd_subset.corr()
    Out[64]:
    playtime_average_player playtime_average_friends playtime_average_friends_max playtime_forever_player playtime_forever_friends game_count_player friend_count group_count_player multiplayer_fraction_player spending_player
    playtime_average_player 1.0000 0.2397 0.2921 0.4708 0.1079 0.2020 0.2837 0.0943 -0.0067 0.2178
    playtime_average_friends 0.2397 1.0000 0.8808 0.1173 0.4190 0.0080 0.1089 0.0156 0.0165 0.0076
    playtime_average_friends_max 0.2921 0.8808 1.0000 0.2258 0.4049 0.0459 0.3261 0.1020 0.0066 0.0456
    playtime_forever_player 0.4708 0.1173 0.2258 1.0000 0.2875 0.3778 0.5178 0.2750 -0.0204 0.3992
    playtime_forever_friends 0.1079 0.4190 0.4049 0.2875 1.0000 0.0822 0.2282 0.1032 -0.0253 0.0834
    game_count_player 0.2020 0.0080 0.0459 0.3778 0.0822 1.0000 0.1962 0.0913 -0.1228 0.9816
    friend_count 0.2837 0.1089 0.3261 0.5178 0.2282 0.1962 1.0000 0.4267 0.0203 0.1994
    group_count_player 0.0943 0.0156 0.1020 0.2750 0.1032 0.0913 0.4267 1.0000 0.0009 0.0886
    multiplayer_fraction_player -0.0067 0.0165 0.0066 -0.0204 -0.0253 -0.1228 0.0203 0.0009 1.0000 -0.1114
    spending_player 0.2178 0.0076 0.0456 0.3992 0.0834 0.9816 0.1994 0.0886 -0.1114 1.0000

    Hypothesis 1: The more friends I have, the more I play

    • Heavy gamers have more friends than regular gamers: average = 40 vs 7 and median 20 vs 1
    • There is a positive, but weak correlation of 0.3 between playtime and the number of friends
    • However, among heavy gamers, the friend count does not correlate with playtime
    • Takeaway: Heavy gamers generally have many friends on the network. But the number of friends does not correlate with their behaviour at the margin.
    In [65]:
    # moments for heavy gamers
user_summary_pd_subset_addicted.friend_count.fillna(0).describe()
    Out[65]:
    count   17,537.0000
mean        40.0153
std         52.4297
min          0.0000
25%          6.0000
50%         22.0000
75%         52.0000
max        674.0000
Name: friend_count, dtype: float64
    In [66]:
    # moments for all gamers
user_summary_pd_subset.friend_count.fillna(0).describe()
    Out[66]:
    count   1,665,521.0000
mean            7.1880
std            19.6804
min             0.0000
25%             0.0000
50%             1.0000
75%             5.0000
max           674.0000
Name: friend_count, dtype: float64
    In [69]:
    # Jointplot
g0 = sns.jointplot("friend_count", "playtime_average_player",  data=user_summary_pd_subset.fillna(0),
                  kind="reg", truncate=False,
                  color="r", height=7, ylim = (0,25), xlim = (0,500), marginal_kws=dict(kde=False),
                  joint_kws = {'scatter_kws':dict(alpha=0.2)})
g0.annotate(stats.pearsonr) 
g0.set_axis_labels('Friend count', 'Average playtime per day', fontsize=20)

g1 = sns.jointplot("friend_count", "playtime_average_player",  data=user_summary_pd_subset_addicted.fillna(0),
                  kind="reg", truncate=False, 
                  color="r", height=7, ylim = (0,25), xlim = (0,500), marginal_kws=dict(kde=False), 
                  joint_kws = {'scatter_kws':dict(alpha=0.2)})
g1.annotate(stats.pearsonr) 
g1.set_axis_labels('Friend count', 'Average playtime per day', fontsize=20)

fig = plt.figure(figsize=(20,10))
gs = gridspec.GridSpec(1, 2)

mg0 = SeabornFig2Grid(g0, fig, gs[0])
mg1 = SeabornFig2Grid(g1, fig, gs[1])

fig.suptitle('          All gamers                                                ' + \
             'Heavy gamers (top 1%) ', fontsize=30)


gs.tight_layout(fig)

plt.show()

    Hypothesis 2: The more my friends play...

    • Heavy gamers' friends (if active) play more than regular gamers' friends (~2h vs 1h per day)
    • The same holds for the median friend (1h45min vs 10min)
    • For heavy gamers, playtime correlation with friends is only 0.03, vs 0.24 for the regular gamer
    • Takeaway: Friends seem to be an influence, but less so at the margin.
    In [70]:
    # moments for heavy gamers
user_summary_pd_subset_addicted.playtime_average_friends.describe()
    Out[70]:
    count   8,341
mean        2
std         2
min         0
25%         0
50%         2
75%         3
max        22
Name: playtime_average_friends, dtype: float64
    In [71]:
    # moments for all gamers
user_summary_pd_subset.playtime_average_friends.describe()
    Out[71]:
    count   217,832
mean          1
std           2
min           0
25%           0
50%           0
75%           1
max          25
Name: playtime_average_friends, dtype: float64
    In [72]:
    # Jointplot
g0 = sns.jointplot("playtime_average_friends", "playtime_average_player", data=user_summary_pd_subset,
                  kind="reg", truncate=False,
                  color="m", height=7, ylim = (0,25), xlim = (0,25), joint_kws = {'scatter_kws':dict(alpha=0.2)})
g0.annotate(stats.pearsonr) 
g0.set_axis_labels('Friends\' average playtime per day', 'Average playtime per day', fontsize=20)

g1 = sns.jointplot("playtime_average_friends", "playtime_average_player", data=user_summary_pd_subset_addicted,
                  kind="reg", truncate=False,
                  color="m", height=7, ylim = (0,25), xlim = (0,25), joint_kws = {'scatter_kws':dict(alpha=0.2)})
g1.annotate(stats.pearsonr) 
g1.set_axis_labels('Friends\' average playtime per day', 'Average playtime per day', fontsize=20)

fig = plt.figure(figsize=(20,10))
gs = gridspec.GridSpec(1, 2)

mg0 = SeabornFig2Grid(g0, fig, gs[0])
mg1 = SeabornFig2Grid(g1, fig, gs[1])

fig.suptitle('          All gamers                                                ' + \
             'Heavy gamers (top 1%) ', fontsize=30)


gs.tight_layout(fig)

plt.show()

    Hypothesis 3: The more my heaviest gamer friend plays...

    • Correlation of playtime with the playtime of the friend that plays the most (0.29) is stronger than with the playtime of the average friend (0.24)
    • But again, this correlation is weaker for the heavy gamers (0.05)
    • Takeaway: The distribution of behaviour among friends matters. The maximum appears more important than the average.
    In [73]:
    # moments for heavy gamers
user_summary_pd_subset_addicted.playtime_average_friends_max.describe()
    Out[73]:
    count   8,341
mean        4
std         3
min         0
25%         1
50%         3
75%         5
max        24
Name: playtime_average_friends_max, dtype: float64
    In [74]:
    # moments for all gamers
user_summary_pd_subset.playtime_average_friends_max.describe()
    Out[74]:
    count   217,832
mean          1
std           2
min           0
25%           0
50%           0
75%           2
max          25
Name: playtime_average_friends_max, dtype: float64
    In [75]:
    # Jointplot
g0 = sns.jointplot("playtime_average_friends_max", "playtime_average_player", data=user_summary_pd_subset,
                  kind="reg", truncate=False,
                  color="purple", height=7, ylim = (0,25), xlim = (0,25), joint_kws = {'scatter_kws':dict(alpha=0.2)})
g0.annotate(stats.pearsonr) 
g0.set_axis_labels('Heaviest gamer friend\'s average playtime per day', 'Average playtime per day', fontsize=20)


g1 = sns.jointplot("playtime_average_friends_max", "playtime_average_player", data=user_summary_pd_subset_addicted,
                  kind="reg", truncate=False,
                  color="purple", height=7, ylim = (0,25), xlim = (0,25), joint_kws = {'scatter_kws':dict(alpha=0.2)})
g1.annotate(stats.pearsonr) 
g1.set_axis_labels('Heaviest gamer friend\'s average playtime per day', 'Average playtime per day', fontsize=20)

fig = plt.figure(figsize=(20,10))
gs = gridspec.GridSpec(1, 2)

mg0 = SeabornFig2Grid(g0, fig, gs[0])
mg1 = SeabornFig2Grid(g1, fig, gs[1])

fig.suptitle('          All gamers                                                ' + \
             'Heavy gamers (top 1%) ', fontsize=30)


gs.tight_layout(fig)

plt.show()

    Hypothesis 4: The more groups I am part of...

    • Rather than friends, it could be the diversity of groups that keeps users engaged
    • Yet, group membership is uncommon: The median gamer is part of 0 groups vs 1 for heavy gamers
    • Correlation with playtime is very low
    • Takeaway: Groups seem to be less important than friends in understanding game addiction
    In [76]:
    # moments for heavy gamers
user_summary_pd_subset_addicted.fillna(0).group_count_player.describe()
    Out[76]:
    count   17,537
mean         7
std         26
min          0
25%          0
50%          1
75%          6
max      1,511
Name: group_count_player, dtype: float64
    In [77]:
    # moments for all gamers
user_summary_pd_subset.fillna(0).group_count_player.describe()
    Out[77]:
    count   1,665,521
mean            2
std             9
min             0
25%             0
50%             0
75%             1
max         1,511
Name: group_count_player, dtype: float64
    In [78]:
    # Jointplot
g0 = sns.jointplot("group_count_player", "playtime_average_player",  data=user_summary_pd_subset.fillna(0),
                  kind="reg", truncate=False,
                  color="orange", height=7, ylim = (0,25), xlim = (0,500), marginal_kws=dict(kde=False),
                  joint_kws = {'scatter_kws':dict(alpha=0.2)})

g0.annotate(stats.pearsonr) 
g0.set_axis_labels('Group count', 'Average playtime per day', fontsize=20)


g1 = sns.jointplot("group_count_player", "playtime_average_player",  data=user_summary_pd_subset_addicted.fillna(0),
                  kind="reg", truncate=False,
                  color="orange", height=7, ylim = (0,25), xlim = (0,500), marginal_kws=dict(kde=False), 
                  joint_kws = {'scatter_kws':dict(alpha=0.2)})
g1.annotate(stats.pearsonr) 
g1.set_axis_labels('Group count', 'Average playtime per day', fontsize=20)

fig = plt.figure(figsize=(20,10))
gs = gridspec.GridSpec(1, 2)

mg0 = SeabornFig2Grid(g0, fig, gs[0])
mg1 = SeabornFig2Grid(g1, fig, gs[1])

fig.suptitle('          All gamers                                                ' + \
             'Heavy gamers (top 1%) ', fontsize=30)


gs.tight_layout(fig)

plt.show()

    Hypothesis 5: The more multiplayer games I play...

    • It would seem that multiplayer games are more addictive, especially given the correlation of playtime and friends
    • However, there is nearly no correlation between the fraction of multiplayer games played and playtime
    • Takeaway: Regular gamers alike heavy gamers play a mix of single and multiplayer games
    In [79]:
    # Jointplot
g0 = sns.jointplot("multiplayer_fraction_player", "playtime_average_player",  data=user_summary_pd_subset,
                  kind="reg", truncate=False,
                  color="blue", height=7, ylim = (0,25), xlim = (0,1), marginal_kws=dict(kde=False), 
                  joint_kws = {'scatter_kws':dict(alpha=0.2)})

g0.annotate(stats.pearsonr) 
g0.set_axis_labels('Fraction of games played that is multiplayer', 'Average playtime per day', fontsize=20)


g1 = sns.jointplot("multiplayer_fraction_player", "playtime_average_player",  data=user_summary_pd_subset_addicted ,
                  kind="reg", truncate=False,
                  color="blue", height=7, ylim = (0,25), xlim = (0,1), marginal_kws=dict(kde=False), 
                  joint_kws = {'scatter_kws':dict(alpha=0.2)})
g1.annotate(stats.pearsonr) 
g1.set_axis_labels('Fraction of games played that is multiplayer', 'Average playtime per day', fontsize=20)

fig = plt.figure(figsize=(20,10))
gs = gridspec.GridSpec(1, 2)

mg0 = SeabornFig2Grid(g0, fig, gs[0])
mg1 = SeabornFig2Grid(g1, fig, gs[1])

fig.suptitle('          All gamers                                                ' + \
             'Heavy gamers (top 1%) ', fontsize=30)


gs.tight_layout(fig)

plt.show()

    What we learn from correlations

    • Several factors have been found to correlate with playtime:
      • The number of friendships
      • The playtime of friends
      • The maximum playtime among friends
      • The number of group memberships
    • Correlation is not causation: We don't know whether the above factors cause addiction
    • But we do know that these factors form important parts of heavy gamer's gaming experiences and need to be explored to understand addiction
    • We also found that these factors seem to play a lesser role in explaining why heavy gamers play an hour more or less (low correlations at the margin)

    Correlations between playtime and engagement factors are lower for heavy gamers

    In [80]:
    # Set up frame
fig, axes = plt.subplots(nrows=1,ncols=2, figsize = (15, 8), sharey = True)

sns.set(font_scale=1.2, style="white") 

# Generate a custom diverging colormap
cmap = sns.diverging_palette(220, 10, as_cmap=True)

labels = ['Avg playtime', 'Friends\' avg playtime', 'Friends\' max playtime', 'Player\'s total playtime', 'Friends\' total playtime', '# Games', '# Friends', '# Groups', 'Mulitplayer games' ]

# Compute the correlation matrix
corr = user_summary_pd_subset.drop(['spending_player'], axis = 'columns').corr()

# Draw the heatmap with the mask and correct aspect ratio
sns.heatmap(corr, cmap=cmap, center=0, vmax = 0.5, vmin = -0.3, 
            square=True, linewidths=.5, cbar_kws={"shrink": .5},
            xticklabels = labels, yticklabels = labels, ax = axes[0])

# Compute the correlation matrix
corr = user_summary_pd_subset_addicted.drop(['spending_player'], axis = 'columns').corr()

# Draw the heatmap with the mask and correct aspect ratio
sns.heatmap(corr, cmap=cmap, center=0, vmax = 0.5, vmin = -0.3, 
            square=True, linewidths=.5, cbar_kws={"shrink": .5},
            xticklabels = labels, yticklabels = labels, ax = axes[1])

axes[0].set_title('All gamers', fontsize = 20)
axes[1].set_title('Heavy gamers (top 1%)', fontsize = 20)

plt.tight_layout()

plt.show()

    Conclusions

    • Game addiction is rare among steam users (playtime distribution follows a power law)
    • The social factor (friends, groups) seems to be key to increased user engagement
    • To study addiction further, we need more information regarding gamers' interaction with games
    In [11]:
    # Remove implausible outliers
user_summary_pd_addicted = user_summary_pd[user_summary_pd.playtime_average_player < 24]
user_summary_pd_addicted = user_summary_pd_addicted[user_summary_pd_addicted.playtime_average_player > 5]
user_summary_pd_addicted.to_csv(os.path.join(path,'heavy_gamers.csv'))